D RESEARCH BRIEF FALL 2021
D Thanks to Dean’s Circle donors like you, MIT Sloan can continue to do what we do best: innovate solutions to difficult challenges. From groundbreaking faculty research agendas to learning opportunities that allow students to apply classroom knowledge in real-world contexts, your annual fund support provides the funding necessary to make MIT Sloan’s mission a reality. This report showcases one example of faculty research made possible by your Dean’s Circle support. Class of 1922 Career Development Professor Danielle Li PhD ’12 recently published a paper that highlights the role of algorithm design on the hiring process and access to job opportunities for diverse applicants. The abstract and introduction from the paper, Hiring as Exploration, are enclosed in this report, as well as an MIT Sloan press release about the implications of Dr. Li’s work. I hope you will enjoy seeing how your generosity impacts the MIT Sloan community. As you read, please know that we at the school remain grateful for your Dean’s Circle support. With gratitude,
Wendy Connors Executive Director of Development
DANIELLE LI Class of 1922 Career Development Professor; Associate Professor, Technological Innovation, Entrepreneurship, and Strategic Management
Danielle Li is the Class of 1922 Career Development Professor and an Associate Professor at the MIT Sloan School of Management, as well as a Faculty Research Fellow at the National Bureau of Economic Research. Her research interests are in economics of innovation and labor economics, with a focus on how organizations evaluate ideas, projects, and people. Danielle’s work has been published in leading academic journals across a range of fields, including the Quarterly Journal of Economics, Science, and Management Science. In addition, her work has been regularly featured in media outlets such as The Economist, The New York Times, and The Wall Street Journal. She has previously taught at the Harvard Business School and the Kellogg School of Management. She holds an AB in mathematics and the history of science from Harvard College and a PhD in economics from MIT. 2
IDEAS MADE TO MATTER
Exploration-based algorithms can improve hiring quality and diversity by Brian Eastwood | Sep 30, 2020
Why It Matters
Identifying job candidates who stand out from the pack yields a more diverse short list for interviews than traditional resume screening practices.
MIT researchers have developed a new approach for using algorithms in the recruiting process that can help companies draw talent from a more diverse pool of job applicants. The approach yields more than three times as many Black and Hispanic candidates than companies may have considered using traditional resume screening algorithms. The algorithm also generates a set of interviewees that is more likely to receive and accept a job offer, which can help companies streamline the hiring process. A new working paper, “Hiring as Exploration,” details the results.
Firms are increasingly turning to algorithms to help them make hiring decisions. Algorithms hold the promise of saving firms time — they can process thousands of applications much faster than a human recruiter could — and also potentially improving screening decisions by unearthing predictors of applicant performance that humans might miss.
Traditional hiring algorithms look for characteristics of a job applicant that predict future success, based on a historical training dataset of applicants who have been interviewed or hired in the past. This type of approach, known as supervised learning, works well when firms have a lot of data on past applicants, and when the qualities that predict past success continue to predict future success. Yet there are many instances when both these assumptions may not be true. For example, applicants from non-traditional
3
backgrounds may be under-represented in the training dataset, making it more difficult for firms to accurately predict their performance. Moreover, skill demands may change over time: firms hiring workers in 2020 may place more emphasis on an employee’s ability to work effectively in a remote setting.
“Static supervised learning approaches may push firms to replicate what has been successful in the past, and that may reduce opportunities for people with non-traditional backgrounds,” said MIT Sloan associate professor Danielle Li, who conducted the research with PhD candidate Lindsey R. Raymond and Columbia University associate professor Peter Bergman. “You need to treat hiring as a dynamic learning problem in order to learn more about the quality of candidates you know less about, in order to make better hiring decisions in the future.”
Diversity matters, but firms struggle to improve Some research has shown that more diverse teams are better for business. When leadership teams come from a diverse background — not just gender and race, but career path and education — companies report greater likelihood of innovation (according to a Boston Consulting Group survey), higher-than-average earnings (a McKinsey & Company study concluded), and a boost to their stock prices (according to Stanford University research).
In addition, a Weber Shandwick survey found that younger workers are increasingly likely to consider diversity and inclusiveness as an important factor in their job search. This suggests that firms that aren’t committed to improving diversity may have trouble attracting top talent, the survey noted.
We were interested in looking at the way algorithm design impacts access to opportunity. Danielle Li | Associate Professor
Attempts to increase diversity in hiring face challenges when firms use traditional machine learning techniques to screen job applicants. Those algorithms focus on selecting the best workers based on what the firm knows right now, rather than considering the possibility that this may lead it to pass over qualified
4
applicants from non-traditional backgrounds who are under-represented in its historical data. Li and coauthors believe that firms should strike a better balance between hiring from groups with proven track records and taking a chance on applicants from less well-represented groups, in order to learn about their abilities.
“Machine learning is increasingly being used to guide decision making — for credit scores, for who receives medical care, and for who gets hired,” Li said. “We were interested in looking at the way algorithm design impacts access to opportunity.”
Dynamic models find applicants who stand out To address this, Li looked at hiring as a dynamic learning problem, analyzing applicants based on their upside potential or option value. The team’s algorithm assigns what it called an “exploration bonus” to identify candidates whose quality the firm knows the least about (given the firm’s existing data). These candidates might be rare based on their educational background, work history, or demographics, but they all share one thing in common: because the firm knows so little about them, it stands to learn the most from giving them a chance. This is referred to as “hiring as exploration” since, as Li put it, “you never know if you don’t try.”
Li, Raymond, and Bergman then applied the dynamic learning algorithm to a data sample of nearly 90,000 job applications that a single Fortune 500 company received over a span of 40 months. Researchers compared their output (the “exploration-oriented model”) to the output of two types of static learning algorithms — one that never changed and one that was updated after a round of 100 applicants (the “supervised learning models”) — and to the firm’s ultimate interview and hiring decisions. The firm was very selective; it rejected roughly 95% of candidates based on its initial resume screen, and only 10% of the candidates who passed the screening accepted a job offer.
Under an exploration-based algorithm, 25% of candidates selected for an interview were hired, up from 10% using human recruiters.
5
Using the supervised learning models, approximately 2% of the applicants who passed the initial resume screening were Black, and less than 5% were Hispanic. Under the exploration-oriented model, the shares rose to 14% and 10%, respectively.
“It’s not because the algorithms are looking to find them, but because the candidates are more rare and the algorithms are exploring them more,” Li said, noting that the algorithm is only designed to maximize quality, without any preference for gender or ethnic diversity.
“Companies and recruiters don’t have to say anything about their preference on demographics or work history,” she said. “The algorithms provide a hands-off way to decide what kinds of diversity to explore.”
The dynamic learning model demonstrated another benefit. The hiring rate among the candidates selected by the algorithms was 25%, compared to a 10% rate for candidates selected by human recruiters. Such a result would enable firms to schedule fewer interviews to fill a position. Li said it could also steer a firm away from continuing to select a large pool of job candidates unlikely to accept a role because they have competing job offers on the table — such as the practice of recruiting MBA graduates from high-profile programs for consulting or financial services roles.
“No one thinks that there aren’t talented people outside the Ivy League. The question is, how do we find them?” Li said. “We created a tool that allows us to identify people from groups that may have been traditionally neglected.”
Click here to read this article online.
6
fill
Hiring as Exploration∗
Danielle Li MIT & NBER
Lindsey Raymond MIT
Peter Bergman Columbia & NBER
February 25, 2021
Abstract This paper views hiring as a contextual bandit problem: to find the best workers over time, firms must balance “exploitation” (selecting from groups with proven track records) with “exploration” (selecting from under-represented groups to learn about quality). Yet modern hiring algorithms, based on “supervised learning” approaches, are designed solely for exploitation. Instead, we build a resume screening algorithm that values exploration by evaluating candidates according to their statistical upside potential. Using data from professional services recruiting within a Fortune 500 firm, we show that this approach improves the quality (as measured by eventual hiring rates) of candidates selected for an interview, while also increasing demographic diversity, relative to the firm’s existing practices. The same is not true for traditional supervised learning based algorithms, which improve hiring rates but select far fewer Black and Hispanic applicants. In an extension, we show that exploration-based algorithms are also able to learn more effectively about simulated changes in applicant hiring potential over time. Together, our results highlight the importance of incorporating exploration in developing decision-making algorithms that are potentially both more efficient and equitable.
JEL Classifications: D80, J20, M15, M51, O33 Keywords: Machine Learning, Hiring, Supervised Learning, Bandit Problems, Algorithmic Fairness, Diversity.
∗ Correspondence to d li@mit.edu, lraymond@mit.edu, and bergman@tc.columbia.edu. We are grateful to David Autor, Pierre Azoulay, Dan Bjorkegren, Emma Brunskill, Max Cytrynbaum, Eleanor Dillon, Alex Frankel, Bob Gibbons, Nathan Hendren, Max Kasy, Pat Kline, Fiona Murray, Anja Sautmann, Scott Stern, John Van Reenen, Kathryn Shaw, and various seminar participants for helpful comments and suggestions. The content is solely the responsibility of the authors and does not necessarily represent the official views of Columbia University, MIT, or the NBER.
8
Algorithms have been shown to outperform human decision-makers across an expanding range of settings, from medical diagnosis to image recognition to game play.1 Yet the rise of algorithms is not without its critics, who caution that automated approaches may codify existing human biases and allocate fewer resources to those from under-represented groups.2 A key emerging application of machine learning (ML) tools is in hiring, a setting where decisions matter for both firm productivity and individual access to opportunity, and where algorithms are increasingly used at the “top of the funnel,” to screen job applicants for interviews.3 Modern hiring ML typically relies on “supervised learning,” meaning that it forms a model of the relationship between applicant covariates and outcomes in a given training dataset, and then applies this model to predict outcomes for subsequent applicants.4 By systematically analyzing historical examples, these tools can unearth predictive relationships that may be overlooked by human recruiters; indeed, a growing literature has shown that supervised learning algorithms can more effectively identify high quality job candidates than human recruiters.5 Yet because this approach implicitly assumes that past examples extend to future applicants, firms that rely on supervised learning may tend to select from groups with proven track records, raising concerns about access to opportunity for non-traditional applicants.6 This paper is the first to develop and evaluate a new class of applicant screening algorithms, one that explicitly values exploration. Our approach begins with the idea that the hiring process can be thought of as a contextual bandit problem: in looking for the best applicants over time, a firm must balance “exploitation” with “exploration” as it seeks to learn the predictive relationship between applicant covariates (the “context”) and applicant quality (the “reward”). Whereas the 1
For example, see Yala et al. (2019); McKinney (2020); Mullainathan and Obermeyer (2019); Schrittwieser et al. (2019); Russakovsky et al. (2015) 2 A widely publicized example is Amazon’s use of an automated hiring tool that penalized the use of the term “women’s” (for example, “women’s crew team”) on resumes: https://www.reuters.com/article/us-amazon-com-jobsautomation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G. Obermeyer et al. (2019); Datta et al. (2015); Lambrecht and Tucker (2019) document additional examples in the academic literature. For additional surveys of algorithmic fairness, see Barocas and Selbst (2016); Corbett-Davies and Goel (2018); Cowgill and Tucker (2019). For a discussion of broader notions of algorithmic fairness, see Kasy and Abebe (2020); Kleinberg et al. (2016). 3 A recent survey of technology companies indicated that 60% plan on investing in AI-powered recruiting software in 2018, and over 75% of recruiters believe that artificial intelligence will transform hiring practices (Bogen and Rieke, 2018). 4 ML tools can be used in a variety of ways throughout the hiring process but, by far, algorithms are most commonly used in the first stages of the application process to decide which applicants merit further human review (Raghavan et al., 2019). In this paper, we will use the term “hiring ML” to refer primarily to algorithms that help make the initial interview decision, rather than the final offer. For a survey of commercially available hiring ML tools, see Raghavan et al. (2019). 5 See, for instance, Hoffman et al. (2018); Cowgill (2018). 6 For example, Kline and Walters (2020) test for discrimination in hiring practices, which can both be related to the use of algorithms and influence the data available to them. The relationship between existing hiring practices and algorithmic biases is theoretically nuanced; for a discussion, see Rambachan et al. (2020); Rambachan and Roth (2019); Cowgill (2018).
1
9
optimal solution to bandit problems is widely known to incorporate some exploration, supervised learning based algorithms engage only in exploitation because they are designed to solve static prediction problems. By contrast, ML tools that incorporate exploration are designed to solve dynamic prediction problems that involve learning from sequential actions: in the case of hiring, these algorithms value exploration because learning improves future choices. Incorporating exploration into screening technologies may also shift the demographic composition of selected applicants. While exploration in the bandit sense—that is, selecting candidates with whatever covariates there is more uncertainty over—need not be the same as favoring demographic diversity, it is also the case that Black, Hispanic, and female applicants are less likely to be employed in high-income jobs, meaning that they will also appear less often in the historical datasets used to train hiring algorithms. Because data under-representation tends to increase uncertainty, adopting bandit algorithms that value exploration (for the sake of learning) may expand representation even when demographic diversity is not part of their mandate.7 Our paper uses data from a large Fortune 500 firm to study the decision to grant first-round interviews for high-skill positions in consulting, financial analysis, and data science—sectors which offer well-paid jobs with opportunities for career mobility and which have also been criticized for their lack of diversity. In this setting, we provide the first empirical evidence that algorithmic design impacts access opportunity. Relative to human screening decisions, we show that contextual bandit algorithms increase the quality of interview decisions (as measured by hiring yield) while also selecting a more diverse set of applicants. Yet, in the same setting, we also show that this is not the case for traditional supervised learning approaches, which increase quality but at the cost of vastly reducing Black and Hispanic representation. Our results therefore demonstrate the potential of algorithms to improve the hiring process, while cautioning against the idea that they are generically equity or efficiency enhancing. Like many firms in its sector, our data provider is overwhelmed with applications and rejects the vast majority of candidates on the basis of an initial resume screen. Motivated by how ML tools are typically used in the hiring process, our goal is to understand how algorithms can impact this consequential interview decision. In our analysis, we focus on hiring yield as our primary measure of quality. Because recruiting is costly and diverts employees from other productive work, our firm would like to adopt screening tools that improve its ability to identify applicants who will ultimately receive and accept an offer; currently, our firm’s hiring rate among those interviewed is only 10%. 7
This logic is consistent with a growing number of studies focusing on understanding persistent biases in employer beliefs, and how to change them. For example, see Miller (2017); Bohren et al. (2019a,b); Lepage (2020a,b). In these papers and others, small sample experiences with some minority workers (or some other source of biased priors) may lock firms into persistent inaccurate beliefs about the overall quality of minority applicants. In such cases, firms may benefit from exploration-based algorithms that nudge them toward obtaining additional signals of minority quality.
10
2
As such, for most of our analysis, we define an applicant’s quality as her “hiring potential”—that is, her likelihood of being hired, were she to receive an interview.8 We build three different resume screening algorithms—two based on supervised learning, and one based on a contextual bandit approach—and evaluate the candidates that each algorithm selects, relative to the actual interview decisions made by the firm’s human resume screeners. We observe data on an applicant’s demographics (race, gender, and ethnicity), education (institution and degree), and work history (prior firms). Our goal is to maximize the quality of applicants who are selected for an interview; although we will also evaluate their diversity, we do not incorporate any explicit diversity preferences into our algorithm design. Our first algorithm uses a static supervised learning approach (hereafter, “static SL”) based on a logit LASSO model. Our second algorithm (hereafter, “updating SL”) builds on the same baseline model as the static SL model, but updates the training data it uses throughout the test period with the hiring outcomes of the applicants it chooses to interview.9 While this updating process allows the updating SL model to learn about the quality of the applicants it selects, it is myopic in the sense that it does not incorporate the value of this learning into its selection decisions. Our third approach implements an Upper Confidence Bound (hereafter, “UCB”) contextual bandit algorithm: in contrast to the static and updating SL algorithms, which evaluates candidates based on their point estimates of hiring potential, a UCB contextual bandit selects applicants based on the upper bound of the confidence interval associated with those point estimates. That is, there is implicitly an “exploration bonus” that is increasing in the algorithm’s degree of uncertainty about quality. Exploration bonuses will tend to be higher for groups of candidates who are underrepresented in the algorithm’s training data because the model will have less precise estimates for these groups. In our implementation, we allow exploration bonuses to be based on a wide set of applicant covariates: the algorithm can choose to assign higher exploration bonuses on the basis of race or gender, but it is not required to and the algorithm could, instead, focus on other variables such as education or work history. Once candidates are selected, we incorporate their realized hiring outcomes into the training data and update the algorithm for the next period.10 Standard and contextual bandit UCB algorithms have been shown to be optimal in the sense that they 8
Henceforce, this paper will use the terms “quality,” “hiring potential,” and “hiring likelihood” interchangeably, unless otherwise noted. 9 In practice, we can only update the model with data from selected applicants who are actually interviewed (otherwise we would not observe their hiring outcome). See Section 3.2.2 for a more detailed discussion of how this algorithm is updated. 10 Similar to the updating SL approach, we only observe hiring outcomes for applicants who are actually interviewed in practice, we are only able to update the UCB model’s training data with outcomes for the applicants it selects who are also interviewed in practice. See Section 3.2.3 for more discussion.
3
11
asymptotically minimize expected regret11 and have begun to be used in economic applications.12 Ours is the first to apply a contextual bandit in the context of hiring. We have two main sets of results. First, our SL and UCB models differ markedly in the demographic composition of the applicants they select to be interviewed. Implementing a UCB model would more than double the share of interviewed applicants who are Black or Hispanic, from 10% to 24%. The static and updating SL models, however, would both dramatically decrease the combined share of Black and Hispanic applicants to 2% and 5%, respectively. In the case of gender, all algorithms would increase the share of selected applicants who are women, from 35% under human recruiting, to 42%, 40%, and 48%, under static SL, updating SL, and UCB, respectively. This increase in diversity is persistent throughout our sample. Our second set of results shows that, despite differences in their demographic profiles, all of our ML models substantial increase hiring yield relative to human recruiters. We note that assessing quality differences between human and ML models is more difficult than assessing diversity because we face a sample selection problem, also known in the ML literature as a “selective labels” problem:13 although we observe demographics for all applicants, we only observe “hiring potential” for applicants who are interviewed. To address this, we take three complementary approaches, all of which consistently show that ML models select candidates with greater hiring potential than human recruiters. We describe these approaches in more detail below. First, we focus on the sample of interviewed candidates for whom we directly observe hiring outcomes. Within this sample, we ask whether applicants preferred by our ML models have a higher likelihood of being hired than applicants preferred by a human recruiter. In order to define a “human score” that proxies for recruiter preferences, we train a fourth algorithm (a supervised learning model similar to our static SL) to predict human interview decisions rather than hiring likelihood (hereafter, “human SL”). We then examine which scores are best able to identify applicants who are hired. We find that, across all of our ML models, applicants with high scores are much more likely to be hired than those with low scores. In contrast, there is almost no relationship between an applicant’s propensity to be selected by a human, and his or her eventual hiring outcome; if anything, this relationship is negative. Our second approach estimates hiring potential for the full sample of applicants. A concern with restricting our analysis to those who are interviewed is that we may overstate the relative accuracy 11
Lai and Robbins (1985); Abbasi-Yadkori et al. (2019); Li et al. (2017) prove regret bounds for several different UCB algorithmns. We follow the approach in Li et al. (2017) that extends the contextual bandit UCB for binary outcomes. See Section 2.1 and 3 for a detailed discussion. 12 For example, see Currie and MacLeod (2020); Stefano Caria and Teytelboym (2020); Kasy and Sautmann (2019); Bergemann and Valimaki (2006); Athey and Wager (2019); Krishnamurthy and Athey (2020); Zhou et al. (2018); Dimakopoulou et al. (2018a). 13 See, for instance, Lakkaraju et al. (2017); Kleinberg et al. (2018a).
12
4
of our ML models if human recruiters add value by making sure particularly weak candidates are never interviewed. To address this concern, we use an inverse propensity score weighting approach to recover an estimate of the mean hiring likelihood among all applicants selected by each of our ML models using information on the outcomes of interviewed applicants with similar covariates. We continue to find that ML models select applicants with higher predicted hiring likelihoods, relative to those selected by humans: average hiring rates among those selected by the UCB, updating SL, and static SL models are 33%, 35%, and 24%, respectively, compared with the observed 10% among observed recruiter decisions. These results suggest that, by adopting an ML approach, the firm could hire the same number of people while conducting fewer interviews. Finally, we use an IV strategy to explore the robustness of our conclusions to the possibility of selection on unobservables. So far, our approaches have either ignored sample selection or have assumed that selection operates on observables only. While there is relatively little scope for selection on unobservables in our setting (recruiters make interview decisions on the basis of resume review only), we verify this assumption using variation from random assignment to initial screeners (who vary in their leniency to grant interviews). In particular, we show that applicants selected by stringent screeners (and therefore subject to a higher bar) have no better outcomes than those selected by more lax screeners: this suggests that humans are not positively screening candidates in their interview decisions. We use this same IV variation to show that firms can improve their current interview practices by following ML recommendations on the margin. Specifically, we estimate the hiring outcomes of instrument compliers, those who would be interviewed only if they are lucky enough to be assigned to a lenient screener. We show that, among these marginal candidates, those with high UCB scores have better hiring outcomes and are also more likely to be Black, Hispanic, or female. This indicates that following UCB recommendations on the margin would increase both the hiring yield and the demographic diversity of selected interviewees. In contrast, following the same logic, we show that following SL recommendations on the margin would generate similar increases in hiring yield but decrease minority representation. These approaches, each based on different assumptions, all yield the same conclusion: ML models increase quality relative to human recruiters, but supervised learning models may do so at the cost of decreased diversity. An alternative explanation for our findings so far is that firms care about on the job performance and recruiters may therefore sacrifice hiring likelihood in order to interview candidates who would perform better in their roles if hired. Our ability to address this concern is unfortunately limited by data availability: we observe job performance ratings for very few employees in our training period, making it impossible to train a model to predict on the job performance. We show, however, that 13
5
our ML models (trained to maximize hiring likelihood) appear more positively correlated with on the job performance than a model trained to mimic the choices of human recruiters. This suggests that it is unlikely that our results can be explained by human recruiters successfully trading off hiring likelihood to maximize other dimensions of quality, insofar as they can be captured by performance ratings or promotions. Together, our main findings show that there need not be an equity-efficiency tradeoff when it comes to expanding diversity in the workplace. Specifically, firms’ current recruiting practices appear to be far from the Pareto frontier, leaving substantial scope for new ML tools to improve both hiring rates and demographic representation. Even though our UCB algorithm places no value on diversity in and of itself, incorporating exploration in our setting would lead our firm to interview twice as many under-represented minorities while more than doubling its predicted hiring yield. At the same time, we emphasize that our SL models lead to similar increases in hiring yield, but at the cost of drastically reducing the number of Black and Hispanic applicants who are interviewed. This divergence in demographic representation between our SL and UCB results demonstrates the importance of algorithmic design for shaping access to labor market opportunities. In addition, we explore two extensions. First, we examine algorithmic learning over time. Our test data cover a relatively short time period, 2018-2019Q1, so that there is relatively limited scope for the relationship between applicant covariates and hiring potential to evolve. In practice, however, this can change substantially over time, both at the aggregate level—due, for instance, to the increasing share of women and minorities with STEM degrees—or at the organizational level—such as if firms improve their ability to attract and retain minority talent. To examine how different types of hiring ML adapt to changes in quality, we conduct simulations in which the hiring potential of one group of candidates substantially changes during our test period. Our results indicate that the value of exploration is higher in cases when the quality of traditionally under-represented candidates is changing. In a second extension, we explore the impact of blinding the models to demographic variables. Our baseline ML models all use demographic variables—race and gender—as inputs, meaning that they engage in “disparate treatment,” a legal gray area.14 To examine the extent to which our results rely on these variables, we estimate a new model in which we remove demographic variables as explicit inputs. We show that this model can achieve similar improvements in hiring yield, but with more modest increases in share of under-represented minorities who are selected. In our data, we see a greater increase in Asian representation because, despite making up the majority of our applicant sample, these candidates are more heterogeneous on other dimensions (such as education 14
14
For a detailed discussion of the legal issues involved in algorithmic decision-making, see Kleinberg et al. (2018b).
6
and geography) and therefore receive larger “exploration bonuses” in the absence of information about race. The remainder of the paper is organized as follows. Section 1 discusses our firm’s hiring practices and its data. Section 2 presents the firm’s interview decision as a contextual bandit problem and outlines how algorithmic interview rules would operate in our setting. Section 3 discuss how we explicitly construct and validate our algorithms. We present our main results on diversity and quality in Section 4, while Sections 5 and 6 discuss our learning and demographics-blinding extensions, respectively.
To read this paper in its entirety online, click here.
7
15
D
D MIT Sloan School of Management Office of External Relations 77 Massachusetts Avenue, E60-200 Cambridge, MA 02139 mitsloan.mit.edu